Researchers introduced DepthBench, a controlled benchmark for studying computational depth in transformers. It varies width-to-depth ratios while keeping model size and training fixed. Standard Pre-LN and most norm-based variants show little benefit as models grow deeper. HC and Full AttnRes designs consistently improve performance even in deep, narrow models. Gains appear in both pre-training loss and domain-specific tasks. Layer-level analysis links these gains to better use of additional layers. The work identifies residual connection design as key to turning architectural depth into effective computation.